From 864d22164cc2a5c563067d96046fa266aeafb94f Mon Sep 17 00:00:00 2001 From: Nick Cao Date: Tue, 16 Jun 2026 10:48:01 -0400 Subject: [PATCH] [Bugfix] Fix DFlash prefix cache corruption due to missing lookahead block (#4449) Signed-off-by: Nick Cao --- vllm_omni/worker/gpu_ar_model_runner.py | 5 +---- 1 file changed, 1 insertion(+), 4 deletions(-) diff --git a/vllm_omni/worker/gpu_ar_model_runner.py b/vllm_omni/worker/gpu_ar_model_runner.py index d29af7314..e55efd3a1 100644 --- a/vllm_omni/worker/gpu_ar_model_runner.py +++ b/vllm_omni/worker/gpu_ar_model_runner.py @@ -1052,10 +1052,7 @@ class GPUARModelRunner(OmniGPUModelRunner, OmniConnectorModelRunnerMixin): spec_config = self.speculative_config propose_drafts_after_bookkeeping = False if spec_config is not None: - input_fits_in_drafter = spec_decode_common_attn_metadata is not None and ( - spec_decode_common_attn_metadata.max_seq_len + self.num_spec_tokens - <= self.effective_drafter_max_model_len - ) + input_fits_in_drafter = self._input_fits_in_drafter(spec_decode_common_attn_metadata) use_gpu_toks = ( spec_config.use_eagle() or spec_config.uses_draft_model() or spec_config.uses_extract_hidden_states() ) and not spec_config.disable_padded_drafter_batch